11 Segmentation and Visualization of Drusen …
297
into account. We first smoothen the OCT images with bilateral filtering [45] and the
location of the RNFL was estimated by detecting the margin of the vitreous with
a threshold. The initial estimate of the RPE layer was obtained through the highly
reflective and locally connected pixels spatially located below the RNFL. We applied
a morphological opening operation (erosion followed by dilation) with a kernel consisting of a disk of 2 µm radius was performed on this initial estimate in order to
remove small isolated regions from the RPE estimate that are mainly artifacts due to
noise present in SD-OCT images, without removing possible drusen larger than this
considered size (2 µm). We then adopted a maximum axial thickness constrain of
20 µm for the RPE estimate, so that the influence of bright pixels wrongly detected
as part of the RPE but that may be mainly due to other RPE abnormalities such as
GA is minimal. Meanwhile, A-scans in the foreground of this estimate that had a
larger number of pixels than this threshold were removed from the initial estimation
and their RPE location was determined by bilinear interpolation. We estimated two
versions of the RPE: (1) the potentially unhealthy (abnormal) RPE in which drusen
may be present was obtained by bilinear interpolation of the initial estimate, and (2)
a healthy (normal) and drusen-free version of the RPE layer was obtained by fitting
the estimated layer with a 3rd polynomial, an operation that would “smooth-out”
any drusen. The areas located between the fitted normal and interpolated RPE layers
were marked as drusen.
The baseline of the projection region used for the RSVP generation was the fitted
lower boundary of the normal RPE layer, while the top boundary of the projection
region was determined by displacing the fitted normal RPE layer anteriorly the same
distance as the largest drusen peak found in the cube (Fig. 11.11). This selected
sub-volume of the OCT excludes structures in the retina that could interrupt with
visualization of drusen, especially the RNFL and choroid (Fig. 11.30).
11.2.2.3 Filling in the Dark Regions of Drusen
For each A-scan of the SD-OCT images, we obtained the maximum intensity pixel
in the interpolated RPE layer and replaced the values of the pixels underneath it
with this maximum intensity value. After the filling process, dark areas of drusen
become bright. Some dark regions of the RPE also become brighter, but the change in
intensity of RPE was minimal compared to that of drusen, since in the RSVP images
we only considered the projection of the pixels between a narrow region using the
fitted drusen-absent RPE as baseline (as shown in Fig. 11.11).
11.2.2.4 Algorithm Evaluation
To effectively evaluate the RSVP approach to drusen segmentation, 46 3D SD-OCT
retinal images from eight patients were analyzed. Each of the 3D OCT images set was
acquired over a 6 × 6 mm area (corresponding to 512 × 128 pixels) with a 1024-pixel
axial resolution on a commercial SD-OCT device (CirrusOCT; Carl Zeiss Meditec,
297
into account. We first smoothen the OCT images with bilateral filtering [45] and the
location of the RNFL was estimated by detecting the margin of the vitreous with
a threshold. The initial estimate of the RPE layer was obtained through the highly
reflective and locally connected pixels spatially located below the RNFL. We applied
a morphological opening operation (erosion followed by dilation) with a kernel consisting of a disk of 2 µm radius was performed on this initial estimate in order to
remove small isolated regions from the RPE estimate that are mainly artifacts due to
noise present in SD-OCT images, without removing possible drusen larger than this
considered size (2 µm). We then adopted a maximum axial thickness constrain of
20 µm for the RPE estimate, so that the influence of bright pixels wrongly detected
as part of the RPE but that may be mainly due to other RPE abnormalities such as
GA is minimal. Meanwhile, A-scans in the foreground of this estimate that had a
larger number of pixels than this threshold were removed from the initial estimation
and their RPE location was determined by bilinear interpolation. We estimated two
versions of the RPE: (1) the potentially unhealthy (abnormal) RPE in which drusen
may be present was obtained by bilinear interpolation of the initial estimate, and (2)
a healthy (normal) and drusen-free version of the RPE layer was obtained by fitting
the estimated layer with a 3rd polynomial, an operation that would “smooth-out”
any drusen. The areas located between the fitted normal and interpolated RPE layers
were marked as drusen.
The baseline of the projection region used for the RSVP generation was the fitted
lower boundary of the normal RPE layer, while the top boundary of the projection
region was determined by displacing the fitted normal RPE layer anteriorly the same
distance as the largest drusen peak found in the cube (Fig. 11.11). This selected
sub-volume of the OCT excludes structures in the retina that could interrupt with
visualization of drusen, especially the RNFL and choroid (Fig. 11.30).
11.2.2.3 Filling in the Dark Regions of Drusen
For each A-scan of the SD-OCT images, we obtained the maximum intensity pixel
in the interpolated RPE layer and replaced the values of the pixels underneath it
with this maximum intensity value. After the filling process, dark areas of drusen
become bright. Some dark regions of the RPE also become brighter, but the change in
intensity of RPE was minimal compared to that of drusen, since in the RSVP images
we only considered the projection of the pixels between a narrow region using the
fitted drusen-absent RPE as baseline (as shown in Fig. 11.11).
11.2.2.4 Algorithm Evaluation
To effectively evaluate the RSVP approach to drusen segmentation, 46 3D SD-OCT
retinal images from eight patients were analyzed. Each of the 3D OCT images set was
acquired over a 6 × 6 mm area (corresponding to 512 × 128 pixels) with a 1024-pixel
axial resolution on a commercial SD-OCT device (CirrusOCT; Carl Zeiss Meditec,